Re: [RFC 0/5] sched: Add CPU rate caps

[Peter Williams <pwil3058@bigpond.net.au>]

Balbir Singh wrote:
> On Fri, May 26, 2006 at 02:20:21PM +1000, Peter Williams wrote:
>> These patches implement CPU usage rate limits for tasks.
>>
>> Although the rlimit mechanism already has a CPU usage limit (RLIMIT_CPU)
>> it is a total usage limit and therefore (to my mind) not very useful.
>> These patches provide an alternative whereby the (recent) average CPU
>> usage rate of a task can be limited to a (per task) specified proportion
>> of a single CPU's capacity.  The limits are specified in parts per
>> thousand and come in two varieties -- hard and soft.  The difference
>> between the two is that the system tries to enforce hard caps regardless
>> of the other demand for CPU resources but allows soft caps to be
>> exceeded if there are spare CPU resources available.  By default, tasks
>> will have both caps set to 1000 (i.e. no limit) but newly forked tasks
>> will inherit any caps that have been imposed on their parent from the
>> parent.  The mimimim soft cap allowed is 0 (which effectively puts the
>> task in the background) and the minimim hard cap allowed is 1.
>>
>> Care has been taken to minimize the overhead inflicted on tasks that
>> have no caps and my tests using kernbench indicate that it is hidden in
>> the noise.
>>
>> Note:
>>
>> The first patch in this series fixes some problems with priority
>> inheritance that are present in 2.6.17-rc4-mm3 but will be fixed in
>> the next -mm kernel.
>>
> 
> 1000 sounds like a course number. A good estimate for the user setting
> these limits would be percentage or better yet let the user decide on the
> parts. For example, the user could divide the available CPU's capacity
> to 2000 parts and ask for 200 parts or divide into 100 parts and as for 10
> parts. The default capacity can be 100 or 1000 parts. May be the part
> setting could be a system tunable.
> 
> I would also prefer making the capacity defined as the a specified portion
> of the capacity of all CPU's. This would make the behaviour more predictable.

The meaning of a cap would change every time you took a CPU off/on line. 
  This makes the behaviour less predictable not more predictable (at 
least in my opinion).  You also have the possibility of a cap being 
larger than the capacity of a single CPU which doesn't make sense when 
capping at the task level.

However, if you still preferred that interface, it could be implemented 
as a wrapper around these functionalities out in user space or inside a 
resource management component.

> 
> Consider a task "T" which has 10 percent of a single CPU's capacity as hard
> limit. If it migrated to another CPU, would the new CPU also make 10% of its
> capacity available "T".
> 
> What is the interval over which the 10% is tracked? Does the task that crosses
> its hard limit get killed? If not, When does a task which has exceeded its
> hard-limit get a new lease of another 10% to use?
> 
> I guess I should move on to reading the code for this feature now :-)

I look forward to your comments.

Peter
-- 
Peter Williams                                   pwil3058@bigpond.net.au

"Learning, n. The kind of ignorance distinguishing the studious."
  -- Ambrose Bierce